hypothesis
active
hypothesis:during-pretraining-the-model-learns-distinct-harm-categories-and-during-finetuning-wires-them-into-general-harmful-request-featuresDuring pretraining the model learns distinct harm categories and during finetuning wires them into general harmful-request features.
Suggested by dictionaries trained on pretraining data having few refusal features.
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- Refusal is not a single linear classifier.
Claims (1)
claim
- Refusals.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- The paper's central mechanistic explanation of why narrow fine-tuning causes broad misalignment
- Figure 12 left and right show accuracy on harmful/ethical identification and 9-way classification.
- Supported by further finetuning and causal interventions.
- From Lindsey et al; modeled with co-activation toy.
- Curating pretraining data to emphasize healthy emotional regulation may shape the model's emotional foundations beneficiallyhypothesis0.791Proposed intervention.
- Shows steering is behaviorally targeted: suppresses general persona drift while preserving intended narrow-domain learning
- Working metaphor introducing the semantic anchoring intuition
- SAEs uncover safety-relevant representations that might be monitored or controlled.